Earlier version, kept for the record (2026-10-14). This draft had the panel loop back and change n, prompt, charts or models after seeing results. We dropped that because re-running until satisfied is optional stopping. The current page is index.html.

How do LLMs frame controversial charts?

VADER sentiment of each model's reply to “Interpret this plot.”, repeated per chart. This flowchart represents a self-driving analysis built for this project: each box reads the latest data and works out its own numbers and state. The human check comes before the main run: a panel reassesses the required n from the pilot's variance and increases it if needed, then the design is frozen and the tests run once. Click a box for the full view.

Loading…